I'm just doing a sanity check for myself and others as it relates to constructing and interpreting confidence intervals with logistic regression. Suppose you specify a logistic regression of $y$ on a binary variable $x$ which returns a coefficient estimate $\hat{\beta_{1}}=0.05$, and a standard error $se_{\hat{\beta_{1}}}=0.001$.

Is it appropriate to claim:

1) The upper and lower bound for the 95% confidence interval for $\hat{\beta}$ is given by $\hat{\beta}{\pm}(1.96\times{se_{\hat{\beta}}})$

2) Since we can convert log odds into a probability with $\frac{e^x}{1+e^x}$, does that mean this transformation also applies to the upper and lower bounds above?

My feeling is that it's not a problem that our confidence interval is not symmetric about ${\hat{\beta}}$. (You can demonstrate this by calculating probabilities for extreme values of the log odds).

I'm also of the opinion that with these kinds of estimates even if you have an estimated coefficient that is statistically significant, it doesn't necessarily imply practical significance. For instance, if you had a coin which you knew had a long run probability of heads equal to .5000000005, a large enough number of trials would find statistical significance of unfairness, but in a practical sense you could consider the coin "fair".

My final question (and this is the hard part), is it appropriate to report these kinds of binomial probabilities in terms of the difference from a fair coin? Like, for the example in the beginning, the log odds being 0.05 instead of saying "the probability $P(y=1|x=1)=0.5125$", we instead subtract 0.5 from that value to say something like "x being equal to 1 and not 0 is associated with a change in the probability" ${\Delta{P(y=1)}=0.0125}$, just to illustrate that that the size of the effect of x on y is small even though there might be statistical significance. I apologize if my notation is screwed up. Hopefully you understand what I'm getting at though.

  • $\begingroup$ log odds or log odds ratio? There are different. $\endgroup$
    – user158565
    Commented Oct 23, 2018 at 21:03
  • $\begingroup$ Oh you're right. There are different. silly me. R spits out the log odds for its coefficients. so That's what I'm referring to $\endgroup$ Commented Oct 23, 2018 at 21:20

1 Answer 1

  1. Yes.

  2. Yes. Because $log\left(\frac {e^x}{1+e^x}\right)$ is monotonic increase function.

  3. It is question about how to express your idea, so reporting the difference is OK. But I prefer p=0.5125, because I do not know what p is if you tell me the $ΔP(y=1)=0.0125$. If you tell me p=0.5125 and I want to know the different from 0.5, I can calculate easily.

It is very good to realize that statistical significance of testing the difference does not mean that difference has real meaning.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.